normalizing RNA-seq data to "unique transcript length" instead of "transcript length"

[lmc]

I've clustered expression profiles from 20 experiments for a large group of highly related genes. I have the raw read counts and normalized this data using [total read counts uniquely matching to a gene]/[total counts in experiment]*[length of transcript]. However, because these genes have a large amount of non-unique sequence, I don't think that this method is correct. I'd like to try normalizing the expression data based on [length of unique k-mers within transcript] rather than [length of transcript]. Is there an existing tool that can calculate this?
Thanks in advance any help!

[mrawlins]

I'm not particularly impressed with the RPKM measure either, as it is still biased towards long transcripts (Oshlack and Wakefield, Biology Direct 2009). The method found in this paper (doi:10.1093/bioinformatics/btp692) seems to be a more intelligent way of addressing this issues, though I haven't yet tested it out. I'm not sure that [length of unique k-mers within transcript] will be any better than [length of transcript] at eliminating the bias you think is there.
If you're set on doing it, though, I think you'll have to roll your own script to determine the length of unique k-mers, which may get to be fairly computationally intensive depending on how you go about doing that.

[steven]

I've clustered expression profiles from 20 experiments for a large group of highly related genes. I have the raw read counts and normalized this data using [total read counts uniquely matching to a gene]/[total counts in experiment]*[length of transcript]. However, because these genes have a large amount of non-unique sequence, I don't think that this method is correct. I'd like to try normalizing the expression data based on [length of unique k-mers within transcript] rather than [length of transcript]. Is there an existing tool that can calculate this?
Thanks in advance any help!

Although it may not remove all the biases, subtracting non unique k-mers from the total transcript length before normalization makes sense. You can find some precomputed data in the "mapability" tracks at the UCSC for this purpose.
Another thing: because of biases in the read coverage at the end of the transcripts, it is frequent to disregard the initial and terminal exons and/or the UTRs.
good luck,
s.